Reward uncertainty in AI benchmarks

Researchers argue current AI test leaderboards penalize models for saying 'I don’t know,' pushing them toward confident guessing and more hallucinations. Changing scoring to reward calibrated uncertainty would realign incentives toward trustworthy behavior and better model selection. This reframes hallucinations as partly a measurement problem, not only a training problem. — If evaluation rules drive model behavior, policy and industry standards must target benchmark design to curb hallucinations and improve reliability.